A transformer fault voiceprint recognition method based on multi-dimensional time-frequency features

Abstract

The application provides a transformer fault sound print recognition method based on multi-dimensional time-frequency characteristics, and relates to the technical field of digital power.The application firstly extracts transformer noise through Fourier transform and a frame method, filters corona, fan and environmental interference data;secondly, compares a transformer full-life-cycle sound print library with a current transformer sound print to determine whether an abnormality exists;and thirdly, adjusts multi-dimensional time-frequency characteristic evaluation weights through an entropy weight method to identify transformer winding fault sound prints.Compared with the prior art, the application has the advantages of high recognition accuracy, no need for power-off maintenance, no need for additional detection equipment except vibration sensors, the ability to determine fault types and good winding fault identification effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital power, and particularly relates to a transformer fault voiceprint recognition method based on multi-dimensional time-frequency characteristics. BACKGROUND

[0002] The main transformer is the core equipment of the transformer substation, and undertakes the task of transforming voltage and distributing electric energy. The safe and reliable operation of the main transformer has a positive significance for the safe and stable operation of the power grid. The main transformer is complex in design, especially the converter transformer is in the quality improvement period, and the failure rate is high. If the transformer fails, it will cause power loss, maintenance cost and other problems. The causes of transformer failure are various, and the state acquisition technology of the transformer is relatively low, and the key hidden dangers cannot be warned in real time. Therefore, we need to intervene in the identification of transformer faults to ensure the normal operation of the transformer.

[0003] The identification of transformer faults mainly includes power-off maintenance and online detection. The power-off maintenance method is to observe whether the winding deforms by using the core lifting method after the transformer is powered off, or to judge whether the transformer has faults by using the capacitance test method, short-circuit impedance method and frequency response curve method. However, the above method needs power-off test, and the test is time-consuming and cannot timely find the hidden danger of the transformer fault. The online detection method is to check by using the transformer partial discharge and transformer oil dissolved gas analysis (DGA). However, the above method has high recognition rate for faults of bushings and oil tanks, but poor recognition effect for winding faults.

[0004] Therefore, it is necessary to provide a transformer fault voiceprint recognition method based on multi-dimensional time-frequency characteristics to solve one of the above technical problems. SUMMARY

[0005] To solve one of the above technical problems, the present application provides a transformer fault voiceprint recognition method based on multi-dimensional time-frequency characteristics, which includes a transformer vibration signal acquisition step, a transformer noise data separation step, a transformer voiceprint feature extraction step, a device full life cycle standard voiceprint comparison step, a fault voiceprint time domain analysis step, a fault voiceprint information entropy analysis step and a fault voiceprint defect identification step.

[0006] Specifically, the transformer vibration signal acquisition step: converts the engineering mechanical vibration parameter of the transformer operation into an electrical signal through a vibration sensor, and measures it, so as to obtain the transformer voiceprint B a .

[0007] Specifically, the transformer noise data separation step: converts the transformer voiceprint B a into a frequency domain signal F a , and separates the frequency domain signal F aThe transformer noise data in the vibration sensor are separated to obtain the vibration signal F after separation of the transformer noise data b .

[0008] Specifically, the transformer voiceprint feature extraction step: the vibration signal F is extracted and filtered by the MFCC mel frequency coefficient b to obtain the transformer voiceprint feature M b .

[0009] Specifically, the device full life cycle standard voiceprint comparison step: including full life cycle standard voiceprint acquisition and full life cycle standard voiceprint comparison; the full life cycle standard voiceprint acquisition is to collect the standard voiceprint data of the whole process of the transformer from the factory test to the delivery, and to establish the full life cycle standard voiceprint; the full life cycle standard voiceprint comparison is to compare the transformer voiceprint feature M b with the full life cycle standard voiceprint, and to calculate the distortion degree; if the distortion degree exceeds the distortion threshold, it is judged that the transformer has a fault, and the fault voiceprint is obtained; otherwise, the transformer is normal.

[0010] Specifically, the fault voiceprint time domain analysis step: the fault voiceprint is analyzed in time domain, and the multi-dimensional time-frequency domain feature and the corresponding weight are obtained.

[0011] Specifically, the fault voiceprint information entropy analysis step: the multi-dimensional time-frequency domain feature of the fault voiceprint is analyzed by the entropy weight method, and the fault voiceprint information entropy of the transformer voiceprint multi-dimensional time-frequency joint analysis is obtained.

[0012] Specifically, the fault voiceprint defect identification step: the information entropy features corresponding to various faults are set in advance, and the fault voiceprint information entropy is identified and classified with the information entropy features of various faults, to obtain the fault category of the transformer.

[0013] As a further solution, the frequency monitoring range of the vibration sensor is 5Hz to 30KHz, and a contact type vibration sensor is installed at multiple mechanical connection positions of the transformer to extract the voiceprint, and the extracted transformer voiceprint B a is:

[0014]

[0015] Wherein, n a is the number of vibration sensors installed on the transformer; b si is the transformer mechanical vibration signal collected by different vibration sensors.

[0016] As a further solution, the transformer noise data separation step converts the sound signal in the vibration sensor into a frequency domain signal F a by Fourier transform:

[0017]

[0018] wherein x(t) is a periodic function of t; t is the time of the periodic function; s(t-t) is a window function in Fourier transform; e -2πjkt is a complex function in Fourier transform.

[0019] As a further solution, the separation of transformer noise data adopts a frame separation method, and in the case of overlapping between two frames after frame separation, the number of transformer noise frames G a is:

[0020]

[0021] wherein n c is the total length of transformer noise data; o a is the length of frame separation; c is the overlap rate between two frames of signals; the transformer noise data is extracted by the number of transformer noise frames G a , and the vibration signal F b is obtained.

[0022]

[0023] wherein n b is the upper limit of 50Hz integer multiples in transformer vibration data; f1 and f2 are 1 times and 2 times of 50Hz in transformer vibration data, respectively.

[0024] As a further solution, the MFCC mel-frequency cepstral coefficient is used for transformer voiceprint feature extraction on the vibration signal F b , and the mel conversion frequency M a is:

[0025]

[0026] wherein d is the transformer vibration frequency.

[0027] As a further solution, the transformer voiceprint M b obtained after filtering is:

[0028]

[0029] wherein n d is the number of points of Fourier transform; G bi is the transformer vibration data after different Fourier transforms; and d i is the parameter of different voiceprint extraction filters.

[0030] As a further solution, the standard voiceprint data from the whole process of factory test to return includes the transformer factory test voiceprint, the handover test voiceprint, the transformer normal operation voiceprint, the device abnormal live detection voiceprint, the fault outage detection voiceprint and the fault test voiceprint.

[0031] As a further solution, the multiple standard voiceprint data are constructed into vector data by vector quantization, and the distortion comparison between vectors is realized by overall quantization in vector space; the average distortion rate H of transformer standard voiceprint comparison b is:

[0032]

[0033] wherein, n g is the number of transformer training vector sets; d(x i ,y i ) is the distance between vectors x and y of different training sets.

[0034] As a further solution, the fault voiceprint time domain analysis step selects the sound intensity level, the high frequency energy proportion, the odd-even harmonic amplitude ratio and the spectral component as the multi-dimensional time-frequency domain features.

[0035] As a further solution, the entropy weight method is used for the transformer voiceprint multi-dimensional time-frequency feature analysis, and the fault voiceprint information entropy R(z a ,z b ,z c ,z d ) of the transformer voiceprint multi-dimensional time-frequency joint analysis is obtained.

[0036] R(z a ,z b ,z c ,z d ) = R(z a ) + R(z b ) + R(z c ) + R(z d ) + R(z a |z b |z c |z d )

[0037] wherein, R(z a ), R(z b ), R(z c ), R(z d ) are respectively the information entropy of the transformer voiceprint sound intensity level, the high frequency energy proportion, the odd-even harmonic amplitude ratio and the spectral component. R(z a |z b |z c |z d) is R(z a ), R(z b ), R(z c ), R(z d ) four information entropy cross section.

[0038] Compared with the related art, the transformer fault sound print recognition method based on multi-dimensional time-frequency features has the following beneficial effects:

[0039] The transformer noise is extracted by Fourier transform and frame method first, and corona, fan, and environmental interference data are filtered; secondly, the transformer full life cycle sound print library is compared with the current transformer sound print to determine whether there is an anomaly; and on this basis, the multi-dimensional time-frequency feature evaluation weight is adjusted by entropy weight method, so that the transformer winding fault sound print is identified. BRIEF DESCRIPTION OF DRAWINGS

[0040] Fig. 1 The flowchart of the transformer fault sound print recognition method based on multi-dimensional time-frequency features is provided.

[0041] Fig. 2 The transformer vibration propagation path diagram is provided. DETAILED DESCRIPTION

[0042] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numbers in the drawings represent the same or similar elements. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

[0043] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations.

[0044] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the specific embodiments below. Those skilled in the art should understand that the present disclosure can also be implemented without some specific details. In some examples, methods, means, elements and circuits that are well known to those skilled in the art are not described in detail, in order to highlight the main idea of the present disclosure.

[0045] As Figs. 1-2As shown, the application provides a transformer fault soundprint recognition method based on multi-dimensional time-frequency characteristics, which includes a transformer vibration signal collection step, a transformer noise data separation step, a transformer soundprint feature extraction step, a device full life cycle standard soundprint comparison step, a fault soundprint time domain analysis step, a fault soundprint information entropy analysis step, and a fault soundprint defect identification step.

[0046] Specifically, the transformer vibration signal collection step: through the vibration sensor, the engineering mechanical vibration parameter of the transformer operation is converted into an electrical signal, and it is measured, so as to obtain the transformer soundprint B a .

[0047] Specifically, the transformer noise data separation step: the transformer soundprint B a is converted into a frequency domain signal F a , and the transformer noise data in the frequency domain signal F a is separated through a frame operation to obtain the vibration signal F b after separation of the transformer noise data.

[0048] Specifically, the transformer soundprint feature extraction step: through the MFCC mel frequency cepstrum coefficient, the transformer soundprint feature extraction and filtering of the vibration signal F b are carried out, and the transformer soundprint feature M b is obtained.

[0049] Specifically, the device full life cycle standard soundprint comparison step: including full life cycle standard soundprint collection and full life cycle standard soundprint comparison; the full life cycle standard soundprint collection is to collect the standard soundprint data of the whole process of the transformer from the factory test to the retirement, and to establish the full life cycle standard soundprint; the full life cycle standard soundprint comparison is to compare the transformer soundprint feature M b with the full life cycle standard soundprint, and to calculate the distortion degree; if the distortion degree exceeds the distortion threshold, it is judged that the transformer has a fault, and the fault soundprint is obtained; otherwise, the transformer is in normal operation.

[0050] Specifically, the fault soundprint time domain analysis step: the fault soundprint is analyzed in time domain, and the multi-dimensional time-frequency domain feature and the corresponding weight are obtained.

[0051] Specifically, the fault soundprint information entropy analysis step: through the entropy weight method, the multi-dimensional time-frequency domain feature of the fault soundprint is analyzed, and the fault soundprint information entropy of the transformer soundprint multi-dimensional time-frequency joint analysis is obtained.

[0052] Specifically, the fault soundprint defect identification step: the information entropy features corresponding to various faults are set in advance, and the fault soundprint information entropy is identified and classified with the information entropy features of various faults, and the fault category of the transformer is obtained.

[0053] It should be noted that: such as Fig. 1 As shown, in the transformer acoustic signature feature extraction stage, vibration sensors installed at the mechanical structural connections of the transformer collect sound data. After separating the noise data using a framing method, the transformer acoustic signature features are extracted. In the transformer standard acoustic signature comparison stage, a typical transformer acoustic signature library is formed using transformer handover test and normal operation data. This library is then compared with the current transformer acoustic signature data to determine if there are any anomalies. If anomalies are found, the acoustic signature is identified as a transformer fault. In the transformer fault acoustic signature defect identification stage, the multi-dimensional time-frequency domain of the transformer acoustic signature is first analyzed, and then the information entropy value of each dimension is calculated to form the transformer fault acoustic signature identification result.

[0054] As a further solution, the vibration sensor has a frequency monitoring range of 5Hz to 30kHz, and contact vibration sensors are installed at multiple mechanical connections of the transformer for acoustic fingerprint extraction. The extracted transformer acoustic fingerprint B a for:

[0055]

[0056] Where, n a The number of vibration sensors installed on the transformer; b si Mechanical vibration signals of transformers collected by different vibration sensors.

[0057] It should be noted that the vibration sensor converts the mechanical vibration parameters of the transformer into electrical signals and measures them to obtain the mechanical vibration characteristics of the transformer. The frequency monitoring range of the vibration sensor is 5Hz to 30kHz. To ensure accurate acquisition of the transformer vibration signal, contact vibration sensors are installed at multiple mechanical connections of the transformer for acoustic signature extraction.

[0058] As a further solution, the transformer noise data separation step converts the sound signal from the vibration sensor into a frequency domain signal F using a Fourier transform. a :

[0059]

[0060] Where x(τ) is a periodic function of τ; τ is the time of the periodic function; σ(τ-t) is the window function in the Fourier transform; e -2πjkt It is a complex function in the Fourier transform.

[0061] It should be noted that the signal extracted by the transformer vibration sensor mainly includes: transformer noise, fan noise, corona noise and environmental noise. Among them, the transformer noise is a stable signal with 50Hz integer multiples, and the frequency range is within 2kHz; the fan noise is a full-band signal within 2kHz; the corona noise is a wide-band short-time pulse signal; the environmental noise is a full-band signal from 20Hz to 20kHz.

[0062] As a further solution, the separation of transformer noise data adopts a frame separation method, and in the case of overlapping between the noise signals of two frames after frame separation, the number of transformer noise frames G a is:

[0063]

[0064] Wherein, n c is the total length of the transformer noise data; o a is the length of the frame; c is the overlap rate between two frames of signals; by the number of transformer noise frames G a extracting transformer noise data, the vibration signal F b :

[0065]

[0066] Wherein, n b is the upper limit of 50Hz integer multiples in the transformer vibration data; f1 and f2 are 1 times and 2 times of 50Hz in the transformer vibration data, respectively.

[0067] It should be noted that since the transformer noise data has a fixed frequency and period, the transformer noise is a stable signal with 50Hz integer multiples, and the frequency range is within 2kHz; and other noises are obviously different from it, so we can extract the transformer noise data by determining the number of transformer noise frames G a .

[0068] As a further solution, the MFCC mel-frequency cepstral coefficient extracts the transformer voiceprint feature from the vibration signal F b , and the mel conversion frequency M a is:

[0069]

[0070] Wherein, d is the transformer vibration frequency.

[0071] As a further solution, the transformer voiceprint M b obtained after filtering is:

[0072]

[0073] Wherein, nd is the number of points of Fourier transform; G bi is the transformer vibration data after different Fourier transform; δ i is the different voiceprint extraction filter parameters.

[0074] As a further solution, the standard voiceprint data from the whole process of factory test to retirement includes the transformer factory test voiceprint, the handover test voiceprint, the transformer normal operation voiceprint, the abnormal live detection voiceprint of equipment, the fault outage detection voiceprint and the fault test voiceprint.

[0075] It should be noted that the transformer full life cycle voiceprint refers to the voiceprint data from the whole process of factory test to retirement of the transformer. It includes the transformer factory test voiceprint, the handover test voiceprint, the transformer normal operation voiceprint, the abnormal live detection voiceprint of equipment, the fault outage detection voiceprint and the fault test voiceprint.

[0076] Firstly, the current transformer voiceprint is compared with the factory test voiceprint, the handover test voiceprint and the normal operation voiceprint to judge the distortion degree of the current voiceprint and the transformer normal voiceprint. If the distortion degree of the current voiceprint and the above voiceprints is large, it means that the transformer has a fault. Secondly, it is compared with the abnormal live detection voiceprint of equipment, the fault outage detection voiceprint and the fault test voiceprint to judge the distortion degree of the current voiceprint and the abnormal voiceprint. If the distortion degree of the current voiceprint and the abnormal voiceprint is small, it means that the transformer has a fault.

[0077] As a further solution, a plurality of standard voiceprint data is constructed into vector data by vector quantization, and the distortion comparison between vectors is realized by overall quantization in the vector space; the average distortion rate H of transformer standard voiceprint comparison b is:

[0078]

[0079] wherein, n g is the number of transformer training vector sets; d(x i ,y i ) is the distance between vectors x and y of different training sets.

[0080] It should be noted that vector quantization (VQ) is a signal correlation quantization method based on Shannon theory. The method constructs a plurality of scalar data into vector data, and realizes the distortion comparison between vectors by overall quantization in the vector space.

[0081] In a specific embodiment, the distortion degrees of the current transformer voiceprint and the transformer full life cycle normal and abnormal voiceprints are respectively counted.

[0082]

[0083] In the formula: △f1 is the factory test sound mark; △f2 is the handover test sound mark; △f3 is the normal operation sound mark; △f4 is the equipment abnormal power-on detection sound mark; △f5 is the fault power outage detection sound mark; △f6 is the fault test sound mark.

[0084] As a further solution, the fault acoustic signature time-domain analysis step selects sound intensity level, high-frequency energy ratio, odd-even harmonic amplitude ratio, and spectral components as multi-dimensional time-frequency domain features.

[0085] It should be noted that transformer winding faults mainly include three types: permanent winding deformation, insulation degradation, and insulation damage. Permanent winding deformation includes situations such as tilting, twisting, displacement, collapse, and bulging of the winding; insulation degradation refers to insufficient short-circuit withstand capability and inter-turn insulation aging; insulation damage is the insulation failure that occurs after prolonged winding deformation and insulation degradation. All of these transformer winding faults will cause changes in the transformer's vibration signature. The transformer vibration propagation path is as follows: Fig. 2 As shown. By Fig. 2 It is evident that when the vibration propagation path of the transformer remains unchanged, the vibration signal changes relatively little. However, when a transformer malfunctions, its mechanical state changes, and the vibration signal changes accordingly.

[0086] Because transformers of different voltage levels and from different manufacturers have significantly different acoustic characteristics under different loads, winding deformations and other operating conditions, we selected multi-dimensional time-frequency domain characteristics, such as sound intensity level, high-frequency energy ratio, odd-even harmonic amplitude ratio and spectral components, based on the typical settings of the transformer industry association for analysis, as shown in Table 1.

[0087] Table 1 Multidimensional Time-Frequency Domain Feature Indicators

[0088]

[0089] As a further solution, the entropy weight method is used to perform multi-dimensional time-frequency feature analysis of transformer acoustic signatures, obtaining the fault acoustic signature information entropy R(z) from the multi-dimensional time-frequency joint analysis of transformer acoustic signatures. a ,z b ,z c ,z d ):

[0090] R(z a ,z b ,z c ,z d )=R(z a )+R(z b )+R(z c )+R(z d )+R(z a |zb |z c |z d )

[0091] wherein R(z a ), R(z b ), R(z c ), R(z d ) are respectively transformer voiceprint sound intensity level, high frequency energy proportion, odd-even harmonic amplitude ratio, information entropy of spectral component. a |z b |z c |z d ) is the cross part of R(z a ), R(z b ), R(z c ), R(z d ) four information entropies.

[0092] It should be noted that: the entropy weight method is a system index weight evaluation method. In this method, the discrete degree of the index is judged by the size of the entropy value. If the discrete degree is greater, the entropy value is smaller, and the weight of the index is greater. Therefore, the entropy weight method is used for transformer voiceprint multi-dimensional time-frequency feature analysis in the paper.

[0093] After obtaining the fault voiceprint information entropy, the transformer winding fault type can be identified through artificial classification or machine learning model classification. Artificial classification or machine learning model classification is an existing classification technology, which will not be described here.

[0094] The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A transformer fault soundprint recognition method based on multi-dimensional time-frequency features, characterized in that, The method comprises a transformer vibration signal collection step, a transformer noise data separation step, a transformer voiceprint feature extraction step, a device full life cycle standard voiceprint comparison step, a fault voiceprint time domain analysis step, a fault voiceprint information entropy analysis step and a fault voiceprint defect identification step. The transformer vibration signal collection step: through the vibration sensor, the engineering machinery vibration parameter of the transformer operation is converted into an electric signal, and it is measured, so as to obtain the transformer voiceprint B of the transformer operation a ; Transformer noise data separation step: convert transformer voiceprint B a into a frequency domain signal F a , and separate the transformer noise data in the frequency domain signal F a by a frame separation operation to obtain a vibration signal F b after transformer noise data separation; Transformer voiceprint feature extraction step: through MFCC mel frequency coefficient to vibration signal F b Perform transformer voiceprint feature extraction and filtering to obtain transformer voiceprint feature M b ; The device full life cycle standard voiceprint comparison step comprises a full life cycle standard voiceprint collection and a full life cycle standard voiceprint comparison; the full life cycle standard voiceprint collection is to collect standard voiceprint data of the whole process of the transformer from factory test to retirement and establish a full life cycle standard voiceprint; the full life cycle standard voiceprint comparison is to compare the transformer voiceprint feature M b with the full life cycle standard voiceprint, and calculate the distortion degree; if the distortion degree exceeds the distortion threshold, it is judged that the transformer fails, and a fault voiceprint is obtained; otherwise, the transformer is in normal operation; The fault voiceprint time domain analysis step: performing time domain analysis on the fault voiceprint and obtaining multi-dimensional time-frequency domain features and corresponding weights; The fault voiceprint information entropy analysis step: performing feature analysis on the multi-dimensional time-frequency domain features of the fault voiceprint by an entropy weight method to obtain fault voiceprint information entropy of transformer voiceprint multi-dimensional time-frequency joint analysis; The fault voiceprint defect identification step: setting information entropy features corresponding to various faults in advance, and identifying and classifying the fault voiceprint information entropy and the information entropy features of various faults to obtain the fault category of the transformer.

2. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 1, characterized in that, The frequency monitoring range of the vibration sensor is 5Hz to 30KHz, and the contact vibration sensor is installed at multiple mechanical connections of the transformer to extract the soundprint. The extracted transformer soundprint B a is: Wherein, n a is the number of vibration sensors installed for the transformer; b si is the transformer mechanical vibration signal collected by different vibration sensors.

3. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 1, characterized in that, The transformer noise data separation step converts the sound signal in the vibration sensor into a frequency domain signal F by Fourier transform a : where x(τ) is a periodic function of τ; τ is the time of the periodic function; σ(τ - t) is a window function in the Fourier transform; e -2πjkt is a complex function in the Fourier transform.

4. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 1, characterized in that, The transformer noise data is separated by frame division, and there is an overlap between the noise signals of two frames after frame division, so the number of transformer noise frames G a is: wherein n c is the total length of the transformer noise data; o a is the length of the frame; c is the overlap rate between two frames of signals; By the transformer noise frame number G a Extracting transformer noise data, the vibration signal F b : wherein n b is the upper limit of the integer multiple of 50Hz in the transformer vibration data; f1 and f2 are 1 times and 2 times of 50Hz in the transformer vibration data, respectively.

5. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 1, characterized in that, MFCC mel-frequency cepstral coefficient pair vibration signal F b Performing transformer voiceprint feature extraction, mel conversion frequency M a is: Wherein, d is the frequency of transformer vibration.

6. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 5, characterized in that, The transformer voiceprint M obtained after filtering b is: wherein n d is the number of Fourier transforms; G bi is the transformer vibration data after different Fourier transforms; δ i is different voiceprint extraction filter parameters.

7. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 1, characterized in that, The standard voiceprint data from the factory test to the whole process of returning includes transformer factory test voiceprint, handover test voiceprint, transformer normal operation voiceprint, device abnormal live detection voiceprint, fault power-off detection voiceprint and fault test voiceprint.

8. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 7, characterized in that, The plurality of standard voiceprint data is constructed into vector data by vector quantization, and the distortion comparison between vectors is realized by overall quantization in vector space; the average distortion rate H of transformer standard voiceprint comparison b is: where n g is the number of transformer training vector sets; d(x i ,y i ) is the distance between vectors x and y of different training sets.

9. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 1, characterized in that, The fault voiceprint time domain analysis step selects sound intensity level, high-frequency energy proportion, odd-even harmonic amplitude ratio and spectral components as multi-dimensional time-frequency domain features.

10. The transformer fault soundprint recognition method based on multi-dimensional time-frequency features according to claim 9, characterized in that, Adopting entropy weight method to analyze the multi-dimension time-frequency characteristics of transformer voiceprint, the fault voiceprint information entropy R(z a ,z b ,z c ,z d ) of transformer voiceprint multi-dimension time-frequency joint analysis is obtained. R(z a ,z b ,z c ,z d ) = R(z a ) + R(z b ) + R(z c ) + R(z d ) + R(z a |z b |z c |z d ) wherein R(z a ), R(z b ), R(z c ), R(z d ) are respectively transformer voiceprint sound intensity level, high frequency energy proportion, odd-even harmonic amplitude ratio, information entropy of spectral components, and R(z a |z b |z c |z d ) is the intersection part of R(z a ), R(z b ), R(z c ), R(z d ) four information entropies.

Images